Custom Character Controller in Unity: Part 3 – Analysis of the Physics API

Up until this point I’ve made multiple references to some of the Unity Physics API, but we haven’t really explored it in detail. As an astute reader may have guessed from the title of this entry, that’s what we’ll be doing now. We’ll go through the functions available, analyze some of their associated issues, and ways to overcome them.

As per usual I’ve done my best to avoid doing original research and will be making heavy use of this post from fhhoollm.

Fhhoollolllmm!!

Fhhoollolllmm!!

The Physics API

With many of the functions simply being variations of each other, it shouldn’t take long to go over the details of the Physics Script Reference. I’m not going to bother talking about the methods that have an All variation available, since they are identical other than that the raycast stops immediately at the first contact.

Raycast: Fires a ray in a specified direction for a specified distance (or infinitely far). If an object is contacted, the RaycastHit structure provides useful information about it: where it was contacted, what the normal of the surface was at the contact point, and so on. Because it fires just an infinitely thin ray, this method isn’t particularly use for collision resolution.

CapsuleCastAll: At first glance this seems ideal for usage with a character controller (due to it’s capsule shape), and for the most part it is. It is important to note that as this is a cast it will only detect a collision where the normal of the surface is facing the cast-no backfaces are detected. In addition, the cast does not detect any objects that are within the boundaries of the “capsule” origin of the cast, i.e., it doesn’t detect any objects touching it’s initial position. This is a drawback we will need to overcome if we want it to be a useful tool for our character controller.

CheckCapsule: Right away we have a candidate to solving the problem stated above. CheckCapsule seems to exactly compliment CapsuleCastAll-it will detect all the objects at the initial position of the cast that the CapsuleCast cannot. Unfortunately, it only returns a bool, as opposed to an array of colliders, giving us no information on what objects we actually collided with.

CheckSphere: Same as above, except with a sphere shape.

Linecast: Identical in terms of function to Raycast. Simply a different way of defining the origin, direction, and magnitude of the ray.

OverlapSphere: Now we’re getting somewhere. As far as I can tell, OverlapSphere works exactly as advertised. Bear in mind this note does appear on the docs:

NOTE: Currently this only checks against the bounding volumes of the colliders not against the actual colliders.

…and I really don’t know what this means. I’ve tested it against Box Colliders, Sphere Colliders and Mesh Colliders and it seems to be checking against the actual collider, not just the bounding volume. Note that I am taking bounding volume to mean axis aligned bounding box, and it may mean something different in this case. If not, I’m going to assume it’s a documentation error.

RaycastAll: Same as the Raycast method, except that it does not stop at the first object it contacts.

SphereCastAll: Functions the same as CapsuleCastAll, with the same primary drawback of not detecting objects contained in the sphere defined at the origin of the cast. SphereCast also (like CapsuleCast) does NOT always return the proper normal of the face it collides with. Because it is a sphere that is being cast (rather than an infinitely thin ray in Raycasting) it can collide with the edges of a mesh. When this happens, the hit.normal that is returned is the interpolated value of the normals of the two faces that are joined by the edge. Since CapsuleCasting is just casting with a swept-sphere, it also has the same issue.

In addition to the above tools to detect collisions, Unity also provides a Rigidbody.SweepTestAll method. After testing it, it seems to have identical behavior to the cast methods; faces contained within the collider are not detected by the sweep. I tend to prefer using CapsuleCastAll and SphereCastAll over SweepTest all, as they offer more options (like being able to define your own origin), however SweepTest is useful for box shaped characters, as there is no BoxCast method.

Mesh Colliders

Before we go any further, I want to talk a little bit about mesh colliders. Up until now we’ve focused primarily on the primitive colliders (Box, Sphere, Capsule, etc). However in practice the overwhelming amount of your level’s collision geometry is going to be composed of mesh colliders.

Unlike the primitive colliders, which have their collision representation built from a variety of preset parameters (radius for spheres, height for boxes, and so on), a mesh collider’s collision data is unsurprisingly formed from a 3D mesh. Mesh colliders come in two flavors: Convex and Concave. This article does a terrific job explaining the difference between them.

Since convex hulls must be fully enclosed and Unity limits their size to 255 polygons, they are unideal to be used to represent intricate level geometry. Concave hulls can be of any size, but they come with the drawback of no longer being an enclosed object; instead of being a solid volume, they are essentially just a surface of triangles. This means we can no longer detect if an object is “inside” a concave mesh, since there is no “inside” to check against. This brings us to the problem of phasing. Phasing occurs when a character is moving fast enough (or a wall collider is small enough) that in two frames he travels from one side of the wall to the other, effectively passing through it. Concave mesh colliders amplify this problem by no longer having the ability to detect player collisions occurring “inside” them, making it easy for the player to phase into the mesh.

Controller movement over one frame. His speed is great enough that neither his initial position or final make contact with the thin mesh collider wall in our way

Controller movement over one frame. His speed is great enough that neither his initial or final position make contact with the thin mesh collider wall in his way

Effectively, if we are directly beside a triangle on the surface of a mesh collider with it’s normal facing in the exact inverse direction of our movement vector, the furthest we can move is exactly equal to twice our radius. Considering collision resolution tends to place the character directly flush with the wall, this is a situation that is encountered fairly often. If your character controller is representing a character of about 2 meters (represented as generic units in Unity) high, your radius is typically in the ballpark of 0.5 meters (units). Which means your character can move at most 1 unit per frame. If your game runs at 30 frames per second, you can move at most 30 meters per second, or 108 kilometers per hour. This is pretty damn fast, but if you’re building the latest and greatest Sonic the Hedgehog title it may not be fast enough.

With the controller directly flush with the surface, it cannot move more than twice it's radius or it will phase through the wall

With the controller directly flush with the surface, it cannot move more than twice it’s radius or it will phase through the wall

One solution to this problem is to run your controller’s physics more than once per frame. Alternatively, we can use CapsuleCastAll to check if there are any colliders between our initial and final position every frame. We’ll explore both these options in future articles where we continue to implement the character controller.

Custom Character Controller in Unity: Part 2 – Implementation

Now that we’ve gone over the basics of character controller collision resolution, I’m going to demonstrate how to implement the last presented example (the pushback method) into Unity.

To start off, make sure you have Unity downloaded and installed. For this article I am using Unity 4.3.4f1. (To check your version of Unity, go HelpAbout Unity…) Open an existing project or create a new one for this tutorial. Create a new scene and create a Cube and a Sphere Game Object within it. Although we’ll eventually move on to using a Capsule shape for our controller, we’ll start with a Sphere to keep it simple. Rename the Sphere to Player and the Cube to Wall. Change the Wall’s scaling factor to 6 on each axis. To ease visualization, I also added a blue transparent material to the player and a green transparent material to the wall. Remove the Sphere Collider component from the player.

This sure beats making those dumb diagrams in Photoshop

This sure beats making those dumb diagrams in Photoshop

Create a new C# script and name it SuperCharacterController.cs, to express our dominance as the alpha character controller. Assign this script to our player, and then copy and paste the following code into it:

using UnityEngine;
using System;
using System.Collections.Generic;

public class SuperCharacterController : MonoBehaviour {

 [SerializeField]
 float radius = 0.5f;

 private bool contact;

 // Update is called once per frame
 void Update () {

 contact = false;

 foreach (Collider col in Physics.OverlapSphere(transform.position, radius))
 {
 Vector3 contactPoint = col.ClosestPointOnBounds(transform.position);

 Vector3 v = transform.position - contactPoint;

 transform.position += Vector3.ClampMagnitude(v, Mathf.Clamp(radius - v.magnitude, 0, radius));

 contact = true;
 }
 }

 void OnDrawGizmos()
 {
 Gizmos.color = contact ? Color.cyan : Color.yellow;
 Gizmos.DrawWireSphere(transform.position, radius);
 }
}

…and that’s all, really. Run the project and open the Scene window, while it’s still running. Drag the player around the edges of the wall and attempt to slowly push him into it. You’ll notice the wall resists, and keeps the player flush against it’s edge. So what are we actually doing here?

Physics.OverlapSphere returns an array of Colliders that are contacted by the sphere. It’s a great function in that it doesn’t come with any of the caveats of the other various methods in the Physics class (which we’ll inspect in more detail later). You define an origin and a radius and it gives you the colliders, no frills.

With any collisions detected, we now need to perform resolution. To retrieve the closest point on the surface of the box collider, we use the ClosestPointOnBounds method. We then take a vector that points from the contactPoint to our location. The vector’s magnitude is then clamped and our position is “pushed” out of the collider the proper amount.

You’ll also notice that I implement OnDrawGizmos so that it’s easy to see when the OverlapSphere is colliding with an object.

 Two frames demonstrating the collision being detected, and then resolved

Two frames demonstrating the collision being detected, and then resolved

Fairly simple. Unfortunately our success up until this point has been…an illusion. Create a new class named DebugDraw.cs, and add in the following code.

using UnityEngine;
using System.Collections;

public static class DebugDraw {

 public static void DrawMarker(Vector3 position, float size, Color color, float duration, bool depthTest = true)
 {
 Vector3 line1PosA = position + Vector3.up * size * 0.5f;
 Vector3 line1PosB = position - Vector3.up * size * 0.5f;

 Vector3 line2PosA = position + Vector3.right * size * 0.5f;
 Vector3 line2PosB = position - Vector3.right * size * 0.5f;

 Vector3 line3PosA = position + Vector3.forward * size * 0.5f;
 Vector3 line3PosB = position - Vector3.forward * size * 0.5f;

 Debug.DrawLine(line1PosA, line1PosB, color, duration, depthTest);
 Debug.DrawLine(line2PosA, line2PosB, color, duration, depthTest);
 Debug.DrawLine(line3PosA, line3PosB, color, duration, depthTest);
 }
}

This is a useful helper function of mine that allows us to draw markers in the editor from anywhere in the code (as opposed to just the OnDrawGizmos function). Modify the foreach loop to look like this.

foreach (Collider col in Physics.OverlapSphere(transform.position, radius))
{
Vector3 contactPoint = col.ClosestPointOnBounds(transform.position);

DebugDraw.DrawMarker(contactPoint, 2.0f, Color.red, 0.0f, false);

Vector3 v = transform.position - contactPoint;

transform.position += Vector3.ClampMagnitude(v, Mathf.Clamp(radius - v.magnitude, 0, radius));

contact = true;
}

Run the code, and you’ll notice when a collision happens a large red cross hair is drawn on it’s location. Now, drag the player inside the wall and observe that the marker follows the player. This isn’t necessarily wrong of the ClosestPointOnBounds function, but to match our pushback model from the previous section we really wanted a ClosestPointOnSurfaceOfBoundsOrSomething.

I can't believe this free game engine doesn't do exactly everything I want all the time

I can’t believe this free game engine doesn’t do exactly everything I want all the time

The main issue here is that we cannot properly resolve collisions when our character’s origin is inside a collider, as we do not have a function that will correctly find the nearest point on the surface. For now however, we’re going to move on to the next problem with our current implementation.

Rotate the wall about 20 degrees either way on it’s y axis and then run the scene. You’ll notice nothing seems to work properly anymore. This is because ClosestPointOnBounds returns the closest point on the axis-aligned bounding box, not the object-oriented bounding box.

Axis-aligned bounding bound on the left, with object-oriented on the right

Axis-aligned bounding bound on the left, with object-oriented on the right

You can already imagine how this problem will extend beyond just Box Colliders. Since the function is only capable of returning the axis-aligned bounding box, it clearly will not give us the closest point on the surface if we’re colliding with any other collider type (Sphere, Capsule, Mesh, etc.). Unfortunately, there is no silver bullet for this issue (or not one I’m aware of); we’ll need to implement a separate algorithm for each collider type.

Let’s start with the easiest first: Sphere Colliders. Create a new Sphere game object in the scene. There are a few steps to finding the nearest point on the surface, none of which are too complicated. To know which direction to push the player, we calculate the direction from our position to the Sphere’s centre. Since every point on a Sphere’s surface is the same distance from the origin, we normalize our direction and then multiply it by our radius and our local scale factor.

The following code implements the above. You’ll notice that in addition to the new method I’ve also added in a conditional check to see what kind of collider our OverlapSphere has detected.

using UnityEngine;
using System;
using System.Collections.Generic;

public class SuperCharacterController : MonoBehaviour {

 [SerializeField]
 float radius = 0.5f;

 private bool contact;

 // Update is called once per frame
 void Update () {

 contact = false;

 foreach (Collider col in Physics.OverlapSphere(transform.position, radius))
 {
 Vector3 contactPoint = Vector3.zero;

 if (col is BoxCollider)
 {
 contactPoint = col.ClosestPointOnBounds(transform.position);
 }
 else if (col is SphereCollider)
 {
 contactPoint = ClosestPointOn((SphereCollider)col, transform.position);
 }

 DebugDraw.DrawMarker(contactPoint, 2.0f, Color.red, 0.0f, false);

 Vector3 v = transform.position - contactPoint;

 transform.position += Vector3.ClampMagnitude(v, Mathf.Clamp(radius - v.magnitude, 0, radius));

 contact = true;
 }
 }

 Vector3 ClosestPointOn(SphereCollider collider, Vector3 to)
 {
 Vector3 p;

 p = to - collider.transform.position;
 p.Normalize();

 p *= collider.radius * collider.transform.localScale.x;
 p += collider.transform.position;

 return p;
 }

 void OnDrawGizmos()
 {
 Gizmos.color = contact ? Color.cyan : Color.yellow;
 Gizmos.DrawWireSphere(transform.position, radius);
 }
}

The astute reader may have noticed that this ClosestPointOn method actually returns the closest point on the surface of the Sphere, unlike the ClosestPointOnBounds which returns the closest point within the bounds. This is handy, but we have a few hurdles to jump before we’re able to make much use of this. For now, let’s tackle the second (and final for today) type of Collider we’ll implement: object-oriented bounding boxes.

Image demonstrates how the vector direction between the origin of the sphere and the location of our controller is extrapolated to give us our nearest point

Image demonstrates how the vector direction between the origin of the sphere and the location of our controller is extrapolated to give us our nearest point

Our general approach to this algorithm will be to take the input point and clamp it within the extents of the box. This will give us the same behaviour as the built in ClosestPointOnBounds, except we’ll ensure that it works even if the box has a rotation other than the identity.

The extents of the Box Collider are defined as it’s local size in x, y, and z. In order to clamp our point to the local extents of the Box Collider, we need to transform it’s position from world coordinates to the local coordinates of the Box Collider’s transform. Once we do that, we can clamp the point’s position within the bounds. To get our final point we then transform it back into world coordinates. The final code for the day looks like this.

using UnityEngine;
using System;
using System.Collections.Generic;

public class SuperCharacterController : MonoBehaviour {

 [SerializeField]
 float radius = 0.5f;

 private bool contact;

 // Update is called once per frame
 void Update () {

 contact = false;

 foreach (Collider col in Physics.OverlapSphere(transform.position, radius))
 {
 Vector3 contactPoint = Vector3.zero;

 if (col is BoxCollider)
 {
 contactPoint = ClosestPointOn((BoxCollider)col, transform.position);
 }
 else if (col is SphereCollider)
 {
 contactPoint = ClosestPointOn((SphereCollider)col, transform.position);
 }

 DebugDraw.DrawMarker(contactPoint, 2.0f, Color.red, 0.0f, false);

 Vector3 v = transform.position - contactPoint;

 transform.position += Vector3.ClampMagnitude(v, Mathf.Clamp(radius - v.magnitude, 0, radius));

 contact = true;
 }
 }

 Vector3 ClosestPointOn(BoxCollider collider, Vector3 to)
 {
 if (collider.transform.rotation == Quaternion.identity)
 {
 return collider.ClosestPointOnBounds(to);
 }

 return closestPointOnOBB(collider, to);
 }

 Vector3 ClosestPointOn(SphereCollider collider, Vector3 to)
 {
 Vector3 p;

 p = to - collider.transform.position;
 p.Normalize();

 p *= collider.radius * collider.transform.localScale.x;
 p += collider.transform.position;

 return p;
 }

 Vector3 closestPointOnOBB(BoxCollider collider, Vector3 to)
 {
 // Cache the collider transform
 var ct = collider.transform;

 // Firstly, transform the point into the space of the collider
 var local = ct.InverseTransformPoint(to);

 // Now, shift it to be in the center of the box
 local -= collider.center;

 // Inverse scale it by the colliders scale
 var localNorm =
 new Vector3(
 Mathf.Clamp(local.x, -collider.size.x * 0.5f, collider.size.x * 0.5f),
 Mathf.Clamp(local.y, -collider.size.y * 0.5f, collider.size.y * 0.5f),
 Mathf.Clamp(local.z, -collider.size.z * 0.5f, collider.size.z * 0.5f)
 );

 // Now we undo our transformations
 localNorm += collider.center;

 // Return resulting point
 return ct.TransformPoint(localNorm);
 }

 void OnDrawGizmos()
 {
 Gizmos.color = contact ? Color.cyan : Color.yellow;
 Gizmos.DrawWireSphere(transform.position, radius);
 }
}

You’ll also notice that I made a few changes to the main collision loop, allowing us to call either the axis-aligned or object-oriented ClosestPointOn in the same line. I say “I made a few changes” in a fairly disingenuous sense in that I really mean “I slightly modified the code that I copied and pasted,” as most of the implementation here is taken from fholm’s RPGController package. You can open the RPGCollisions class within it to check out some of the other alterations I made: namely, updating some deprecated code and replacing the matrix multiplications with the more user friendly TransformPoint methods.

Cut me some slack I got like a C minus minus in Linear Algebra I need all the user friendliness I can get.

Cut me some slack I got like a C minus minus in Linear Algebra I need all the user friendliness I can get

This wraps up the first part of our implementation. In future articles I’ll address some of the shortcomings with Unity’s physics API that I’ve alluded to, and begin to outline various components of our ideal Character Controller we will build.

Maybe. Who knows. I’m just making this stuff up as I go.

References

The majority of the code from this article comes from fholm’s RPGController package, specifically the PushBack method from RPGMotor.cs and the closest point methods from RPGCollisions.cs.

Custom Character Controller in Unity: Part 1 – Collision Resolution

After using Unity over the years for various projects, I’ve come to two conclusions: overall it’s a terrific engine that I would recommend to anyone interested in getting into game development, and that it’s built in character controller sucks.  I’ve been working on a custom character controller for a couple weeks and noticed that finding any kind of reference or learning material on the subject is pretty difficult.  So since I couldn’t find anything out there to read…I figured I’d write something instead!  I intend to post a few pieces here outlining what I’ve learned so far and some issues I’ve encountered.  For any actual implementation, I’ll be using the aforementioned Unity game engine.  You can visit their website here and download their latest version here.  I really dig Unity.  It takes care of a lot of the low level under the hood side of game development but still gives you enough freedom to do just about anything.  It also has a really great and active community that is the epitome of helpfulness.

Unfortunately, as previous stated Unity is also home to the world’s worst character controller.  I suppose I should loosely define what a “character controller” is before I build my case against Unity’s.  More or less it’s the code (or class, or whatever) that handles and resolves your character’s collisions in the world.  Unlike boxes and barrels and whatnot that can be taken care of by the rest of the physics engine, characters require special code to behave differently.  However, since we’re doing collision tests, we still need to pick a geometric shape to represent our character.  Most 3D games use a capsule collider.

Capsule collider used to approximate a character's form in Unity

Capsule collider used to approximate a character’s form in Unity

Capsule colliders are great for a wide variety of reasons, but we’ll see that later.  For now, I’m going to go over the basics of character controllers in just 2 dimensions, to keep it simple for now.

CCSetup1

You’ll notice I’ve labelled the axes and x, instead of and y.  This is because I’m going to treat this as a top down view of a three dimensional world, which we’ll eventually move onto.  The character here is seen as a blue circle, as a capsule seen from top view is a circle!  The green rectangle we will treat as a wall.  Ideally, characters cannot walk through walls.  So should the character intersect with it, we’ll want to detect the collision, and properly resolve it.  We’re going to pass over the actual collision detection, i.e., checking to see if the circle is intersecting the box, for two reasons.  One is that Unity has a fair amount of resources (which we’ll go over later) to handle this, and two is that there is a pretty good selection of reference material out there for collision detection.  We’ll focus on getting the colliding object to resolve it’s position properly based on expected behavior.

CCSetup2

Shown above we have the controller attempting to move into the wall.  To prevent this we run a function that performs a sweep test, from the initial position of the controller to the desired position of the movement.  The test detects that a wall is in front of us, and returns the distance.  Using the value we move the controller in the direction of the test the distance the check traveled, placing it directly beside the wall.  (Aside: Unity has several built in functions for this, including Rigidbody.Sweeptest, Physics.SphereCast and Physics.CapsuleCast).

However, this really isn’t the kind of behavior we want.  If we use this method, the character will be immediately halted in any movement he makes if he collides with an object, if even only slightly.  This is undesirable as it doesn’t reflect the way real world objects tend to bounce and slide off each and, more importantly, it would be annoying as hell to play.

CCSetup3

This is a much more desirable behavior.   The initial sweep test is performed in the movement direction for the movement distance.  When the sweeptest contacts the wall, the character is moved directly to it, just like before.  However, this time around we further move the character upwards to make up for the lost movement, which allows it to slide along surfaces.  This is a great example of the desired behavior of the controller, but it isn’t the best way to implement it.  For one, it’s not very efficient: every time you want to move the controller, you need to run this function.  This is fine if you just move it once per frame, but if you plan on doing something different–for whatever reason–you’ll need to rerun the function.  Two, collisions resolution is reliant on character movement direction and distance.  If he just magically finds his way into a solid wall (as character controllers are wont to do) he’s not going to get automatically pushed out.  In practice, I’ve found this method a massive headache.

CCSetup4

Here is that terrifying situation put onto screen.  We see our hero is currently within the walls of the object.  Instead of looking at collision resolution as a response to a movement, like the previous examples, instead we are going to treat it completely independently.  We no longer are concerned with what direction the player moved or how far he moved.  Instead, we will consider only where he is at this moment, and whether his location is a problem or not.  In the above figure, we can see that currently the player is intersecting the wall (he’s inside it!), and therefore his current location is a problem and needs to be rectified.  Since we are no longer resolving collision as a response to movement, we do not know where his previous position was or how far he moved.  All we know is that currently he is stuck inside a wall, and we need to move him out of the wall.  But where to put him?  Just like in previous examples, we should only push him out so that he is just touching the edge of the wall.  We have many locations that are candidates for this…

CCSetup7

Each of the transparent yellow circles indicate a possible position for the character controller that satisfies our goal–to push him out of the wall to a point somewhere on it’s surface.  But which point do we choose?  Simply put, calculate the closest point on the surface of the wall with respect to the controller’s location.

Drawing these diagrams in Photoshop was a huge mistake they take forever

Drawing these diagrams in Photoshop was a huge mistake they take forever

Here we have calculated that the nearest point to our controller’s current location lies to the right of us.  We then move the controller to that point, plus the radius of the controller (shown in red).

This concludes the first part of our epic adventure into the mysterious realm of character controllering.  Next time I’ll start talking about my implementation in Unity, and some of the more complex functions character controllers use.

Acknowledgements and References:

Most of the knowledge here I’ve acquired by exploring two main information sources: a Unity forums post by the user techmage, and reading through/reverse engineering a Unity custom character controller package by user fholm.  I don’t know how to pronounce that either.  Fuhholllme.  Disgusting.  Reminds me of phlegm.  Anyways, you can download his package off his github here, under the RPGController directory.  This is an amazing project overall that I will be exploring the next few sections, and is really terrifically coded.  Apparently he does Unity consulting too if anyone needs that kind of services.  Finally, if anyone has any good reference material for this subject, or anything relating, sharing it here would be a really great way to expand on the topic!